Re: v4.2-rc dcache regression, probably 75a6f82a0d10

[Dominique Martinet]

Al Viro wrote on Sat, Aug 01, 2015:
> And that has turned the check done to an inode that *was* ours at some
> point (i.e. fetching it had been followed by checking that ->d_seq had
> been still valid) into something completely unprotected.  Suppose we
> are in lazy mode and somebody had evicted nd->path.dentry after we'd looked
> it up and before that check.  Sure, its ->d_seq had been bumped by that,
> and we would've failed anyway.  With ECHILD.  Which, unlike ENOTDIR, is
> "repeat in non-lazy mode".

That sounds like a good find, I was looking at how to claim/protect the
entry somehow as well but I just have no idea...

> Folks, could you check if this fixes the problems you are seeing?
> 
> diff --git a/fs/namei.c b/fs/namei.c
> index ae4e4c1..b16c3a7 100644
> --- a/fs/namei.c
> +++ b/fs/namei.c
> @@ -1954,7 +1954,11 @@ OK:
>  				continue;
>  			}
>  		}
> -		if (unlikely(!d_can_lookup(nd->path.dentry)))
> +		if (unlikely(!d_can_lookup(nd->path.dentry))) {
> +			if (nd->flags & LOOKUP_RCU) {
> +				if (unlazy_walk(nd, NULL, 0))
> +					return -ECHILD;
> +			}
>  			return -ENOTDIR;
>  	}
>  }

Unfortunately, still happens for me.

I had to adapt a bit because using an old kernel (4bf46a272), will try
again with a recent master to doublecheck, but I had a break on
the "if (nd->flags & LOOKUP_RCU)" check:
 - sometimes fails without ever hitting the check. I think this fixes
the "ENOTDIR" I had described, but there's at least another way to
fail?
 - When we do hit it, we're into LOOKUP_RCU at this point alright,
unlazy_walk fails and we try again without RCU -- can confirm the
recovery process goes OK (well, that it went OK at least once)


Thanks,
-- 
Dominique
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/